Merge branch 'main' into nicer-app-emails

Author: samejr

apps/supervisor/src/env.ts

@@ -35,8 +35,16 @@ const Env = z.object({
   TRIGGER_DEQUEUE_ENABLED: BoolEnv.default(true),
   TRIGGER_DEQUEUE_INTERVAL_MS: z.coerce.number().int().default(250),
   TRIGGER_DEQUEUE_IDLE_INTERVAL_MS: z.coerce.number().int().default(1000),
-  TRIGGER_DEQUEUE_MAX_RUN_COUNT: z.coerce.number().int().default(10),
-  TRIGGER_DEQUEUE_MAX_CONSUMER_COUNT: z.coerce.number().int().default(1),
+  TRIGGER_DEQUEUE_MAX_RUN_COUNT: z.coerce.number().int().default(1),
+  TRIGGER_DEQUEUE_MIN_CONSUMER_COUNT: z.coerce.number().int().default(1),
+  TRIGGER_DEQUEUE_MAX_CONSUMER_COUNT: z.coerce.number().int().default(10),
+  TRIGGER_DEQUEUE_SCALING_STRATEGY: z.enum(["none", "smooth", "aggressive"]).default("none"),
+  TRIGGER_DEQUEUE_SCALING_UP_COOLDOWN_MS: z.coerce.number().int().default(5000), // 5 seconds
+  TRIGGER_DEQUEUE_SCALING_DOWN_COOLDOWN_MS: z.coerce.number().int().default(30000), // 30 seconds
+  TRIGGER_DEQUEUE_SCALING_TARGET_RATIO: z.coerce.number().default(1.0), // Target ratio of queue items to consumers (1.0 = 1 item per consumer)
+  TRIGGER_DEQUEUE_SCALING_EWMA_ALPHA: z.coerce.number().min(0).max(1).default(0.3), // Smooths queue length measurements (0=historical, 1=current)
+  TRIGGER_DEQUEUE_SCALING_BATCH_WINDOW_MS: z.coerce.number().int().positive().default(1000), // Batch window for metrics processing (ms)
+  TRIGGER_DEQUEUE_SCALING_DAMPING_FACTOR: z.coerce.number().min(0).max(1).default(0.7), // Smooths consumer count changes after EWMA (0=no scaling, 1=immediate)
 
   // Optional services
   TRIGGER_WARM_START_URL: z.string().optional(),
@@ -77,6 +85,10 @@ const Env = z.object({
   KUBERNETES_EPHEMERAL_STORAGE_SIZE_LIMIT: z.string().default("10Gi"),
   KUBERNETES_EPHEMERAL_STORAGE_SIZE_REQUEST: z.string().default("2Gi"),
   KUBERNETES_STRIP_IMAGE_DIGEST: BoolEnv.default(false),
+  KUBERNETES_CPU_REQUEST_MIN_CORES: z.coerce.number().min(0).default(0),
+  KUBERNETES_CPU_REQUEST_RATIO: z.coerce.number().min(0).max(1).default(0.75), // Ratio of CPU limit, so 0.75 = 75% of CPU limit
+  KUBERNETES_MEMORY_REQUEST_MIN_GB: z.coerce.number().min(0).default(0),
+  KUBERNETES_MEMORY_REQUEST_RATIO: z.coerce.number().min(0).max(1).default(1), // Ratio of memory limit, so 1 = 100% of memory limit
 
   // Placement tags settings
   PLACEMENT_TAGS_ENABLED: BoolEnv.default(false),

apps/supervisor/src/index.ts

@@ -128,7 +128,18 @@ class ManagedSupervisor {
       dequeueIdleIntervalMs: env.TRIGGER_DEQUEUE_IDLE_INTERVAL_MS,
       queueConsumerEnabled: env.TRIGGER_DEQUEUE_ENABLED,
       maxRunCount: env.TRIGGER_DEQUEUE_MAX_RUN_COUNT,
-      maxConsumerCount: env.TRIGGER_DEQUEUE_MAX_CONSUMER_COUNT,
+      metricsRegistry: register,
+      scaling: {
+        strategy: env.TRIGGER_DEQUEUE_SCALING_STRATEGY,
+        minConsumerCount: env.TRIGGER_DEQUEUE_MIN_CONSUMER_COUNT,
+        maxConsumerCount: env.TRIGGER_DEQUEUE_MAX_CONSUMER_COUNT,
+        scaleUpCooldownMs: env.TRIGGER_DEQUEUE_SCALING_UP_COOLDOWN_MS,
+        scaleDownCooldownMs: env.TRIGGER_DEQUEUE_SCALING_DOWN_COOLDOWN_MS,
+        targetRatio: env.TRIGGER_DEQUEUE_SCALING_TARGET_RATIO,
+        ewmaAlpha: env.TRIGGER_DEQUEUE_SCALING_EWMA_ALPHA,
+        batchWindowMs: env.TRIGGER_DEQUEUE_SCALING_BATCH_WINDOW_MS,
+        dampingFactor: env.TRIGGER_DEQUEUE_SCALING_DAMPING_FACTOR,
+      },
       runNotificationsEnabled: env.TRIGGER_WORKLOAD_API_ENABLED,
       heartbeatIntervalSeconds: env.TRIGGER_WORKER_HEARTBEAT_INTERVAL_SECONDS,
       sendRunDebugLogs: env.SEND_RUN_DEBUG_LOGS,

apps/supervisor/src/workloadManager/kubernetes.ts

@@ -20,6 +20,12 @@ export class KubernetesWorkloadManager implements WorkloadManager {
   private namespace = env.KUBERNETES_NAMESPACE;
   private placementTagProcessor: PlacementTagProcessor;
 
+  // Resource settings
+  private readonly cpuRequestMinCores = env.KUBERNETES_CPU_REQUEST_MIN_CORES;
+  private readonly cpuRequestRatio = env.KUBERNETES_CPU_REQUEST_RATIO;
+  private readonly memoryRequestMinGb = env.KUBERNETES_MEMORY_REQUEST_MIN_GB;
+  private readonly memoryRequestRatio = env.KUBERNETES_MEMORY_REQUEST_RATIO;
+
   constructor(private opts: WorkloadManagerOptions) {
     this.k8s = createK8sApi();
     this.placementTagProcessor = new PlacementTagProcessor({
@@ -63,6 +69,10 @@ export class KubernetesWorkloadManager implements WorkloadManager {
     return imageRef.substring(0, atIndex);
   }
 
+  private clamp(value: number, min: number, max: number): number {
+    return Math.min(Math.max(value, min), max);
+  }
+
   async create(opts: WorkloadManagerCreateOptions) {
     this.logger.log("[KubernetesWorkloadManager] Creating container", { opts });
 
@@ -295,9 +305,16 @@ export class KubernetesWorkloadManager implements WorkloadManager {
   }
 
   #getResourceRequestsForMachine(preset: MachinePreset): ResourceQuantities {
+    const cpuRequest = preset.cpu * this.cpuRequestRatio;
+    const memoryRequest = preset.memory * this.memoryRequestRatio;
+
+    // Clamp between min and max
+    const clampedCpu = this.clamp(cpuRequest, this.cpuRequestMinCores, preset.cpu);
+    const clampedMemory = this.clamp(memoryRequest, this.memoryRequestMinGb, preset.memory);
+
     return {
-      cpu: `${preset.cpu * 0.75}`,
-      memory: `${preset.memory}G`,
+      cpu: `${clampedCpu}`,
+      memory: `${clampedMemory}G`,
     };
   }